DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Pith reviewed 2026-05-23 04:47 UTC · model grok-4.3
The pith
Pure reinforcement learning on LLMs produces emergent reasoning patterns and outperforms supervised models trained on human demonstrations on verifiable math, coding, and STEM tasks.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation.
Load-bearing premise
That reward signals derived solely from verifiable final answers on training tasks are sufficient to produce generalizable reasoning strategies that transfer to unseen complex problems.
read the original abstract
General reasoning represents a long-standing and formidable challenge in artificial intelligence. Recent breakthroughs, exemplified by large language models (LLMs) and chain-of-thought prompting, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent upon extensive human-annotated demonstrations, and models' capabilities are still insufficient for more complex problems. Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation. Consequently, the trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields, surpassing its counterparts trained via conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims that pure reinforcement learning (RL) can incentivize advanced reasoning capabilities in LLMs without requiring human-labeled reasoning trajectories. The proposed RL framework is said to enable emergent behaviors such as self-reflection, verification, and dynamic strategy adaptation. As a result, the trained model (DeepSeek-R1) achieves superior performance on verifiable tasks including mathematics, coding competitions, and STEM fields compared to models trained via supervised fine-tuning on human demonstrations. The emergent patterns are further claimed to be harnessable for improving smaller models.
Significance. If substantiated with rigorous evidence, the result would be significant for reducing dependence on human-annotated data in scaling LLM reasoning. Demonstrating that outcome-based rewards alone can induce transferable reasoning strategies would challenge current reliance on supervised fine-tuning for complex tasks. However, the abstract supplies no metrics, baselines, training details, or statistical evidence, so the significance cannot be assessed from the provided text; the full manuscript would need to include these to support the claims.
major comments (2)
- [Abstract] Abstract: the central claim that pure RL produces superior performance and emergent reasoning patterns is asserted without any quantitative metrics, baselines (e.g., specific SFT models), training details, or statistical evidence. This absence makes the claim unevaluable and directly undermines assessment of whether outcome-only rewards suffice for generalizable strategies.
- [Abstract] Abstract: the assertion that reward signals from verifiable final answers induce transferable patterns such as self-reflection and dynamic strategy adaptation on unseen problems lacks any supporting isolation experiment or transfer results; the training distribution (math/coding/STEM with easy verification) does not by itself guarantee generalization when problem structure changes.
Simulated Author's Rebuttal
We thank the referee for the constructive comments on the abstract. We agree that the abstract would be strengthened by the inclusion of key quantitative results and will revise it in the next version. We address each major comment below.
read point-by-point responses
-
Referee: [Abstract] Abstract: the central claim that pure RL produces superior performance and emergent reasoning patterns is asserted without any quantitative metrics, baselines (e.g., specific SFT models), training details, or statistical evidence. This absence makes the claim unevaluable and directly undermines assessment of whether outcome-only rewards suffice for generalizable strategies.
Authors: The full manuscript contains extensive experimental sections with quantitative metrics, specific SFT baselines, training details, and performance comparisons on mathematics, coding, and STEM benchmarks. We will revise the abstract to include representative numerical results and baseline references so that the central claims can be evaluated directly from the abstract. revision: yes
-
Referee: [Abstract] Abstract: the assertion that reward signals from verifiable final answers induce transferable patterns such as self-reflection and dynamic strategy adaptation on unseen problems lacks any supporting isolation experiment or transfer results; the training distribution (math/coding/STEM with easy verification) does not by itself guarantee generalization when problem structure changes.
Authors: The manuscript presents analyses of emergent behaviors during RL training and dedicated experiments showing that the resulting reasoning patterns can be used to improve smaller models. While we do not claim the training distribution guarantees generalization to arbitrary problem structures outside the evaluated domains, the reported results include transfer within verifiable tasks. We will add a brief reference to the transfer experiments in the revised abstract. revision: partial
Circularity Check
No circularity in claimed derivation or results
full rationale
The paper reports an empirical RL training procedure on verifiable-outcome tasks (math, coding, STEM) and measures downstream performance plus emergent behaviors such as self-reflection. No equations, uniqueness theorems, fitted parameters renamed as predictions, or self-citation chains are invoked to derive the central claim; the results are presented as experimental outcomes rather than reductions to inputs by construction. The derivation chain is therefore self-contained against external benchmarks.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Reinforcement learning with outcome-based rewards can shape complex sequential behaviors in large neural networks.
Forward citations
Cited by 60 Pith papers
-
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
Noisy expert imitation learning requires exponential samples for offline methods but polynomial for a variant of on-policy distillation under a noise condition.
-
On the Policy Gradient Foundations of Group Relative Policy Optimization: Credit Assignment, Gradient Sparsity, and Rank Collapse
GRPO's group-mean baseline assigns identical advantages to all tokens under output-only rewards, inducing gradient sparsity and an intrinsic rank-2 structure proven from the zero-sum constraint and confirmed by SVD on...
-
Learning to Trigger: Reinforcement Learning at the Large Hadron Collider
RL agent for online LHC trigger threshold tuning improves in-tolerance intervals by 28-56% on Monte Carlo and real CMS data without fine-tuning.
-
Entropy-Gated Latent Recursion
EGLR adds a deterministic layer-recursion axis gated by entropy that is complementary to temperature sampling, raising joint oracle accuracy on MATH-500 from 83.4% to 91.6% for a 3B model.
-
EHRNote-ChatQA: A Benchmark for Evidence-Grounded Multi-Turn Clinical Question Answering over Longitudinal Discharge Summaries
EHRNote-ChatQA is the first benchmark for evidence-grounded multi-turn clinical QA over longitudinal discharge summaries, containing 16,072 medical-expert-verified pairs across eight categories and revealing LLM weakn...
-
HKJudge: A Legal Discourse-Annotated Corpus for Interpreting What Courts Find, How They Reason, and What They Rule
HKJudge is a new ~290k-sentence expert-annotated corpus of Hong Kong criminal judgments with 26 rhetorical roles and 3 sentencing elements, plus benchmarks on classification and extraction tasks.
-
A Pre-Registered Causal Partition of Self-Consistency Elicitation and Reward Design in RLVR
Derives an exact telescoping decomposition of the naive RLVR reward-design estimator into null, elicitation, and reward-design terms on a tabular-GRPO simulator, measures the components across prior strengths, and val...
-
UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing
UltraEP is the first exact-load real-time expert balancer for large-EP MoE training and serving on rack-scale nodes, reaching 94.3% of ideal throughput and 1.49x over no-balancing.
-
Rethinking the Role of Positional Encoding: Sliding-Window Transformers without PE Remain Turing Complete
Sliding-window transformers without positional encodings are Turing complete because the sliding window breaks permutation symmetry and suffices to simulate Post machines via a constant-size histogram state.
-
SVFSearch: A Multimodal Knowledge-Intensive Benchmark for Short-Video Frame Search in the Gaming Vertical Domain
SVFSearch is the first open benchmark for short-video frame search in the Chinese gaming domain, with evaluations showing direct QA at 66.4%, best practical agents at 79.1%, and oracle knowledge at 95.4%.
-
Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination
VLMs fail to detect semantically different image swaps up to 60% of the time despite self-reflective statements, with thinking models more vulnerable and attention analysis showing self-reflection does not increase vi...
-
REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations
REALISTA optimizes continuous combinations of valid editing directions in latent space to produce realistic adversarial prompts that elicit hallucinations more effectively than prior methods, including on large reason...
-
Sieve: Dynamic Expert-Aware PIM Acceleration for Evolving Mixture-of-Experts Models
Sieve dynamically schedules MoE experts across GPU and PIM hardware to handle bimodal token distributions, achieving 1.3x to 1.6x gains in throughput and interactivity over static prior PIM systems on three large models.
-
Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs
Soohak is a 439-problem mathematician-curated benchmark where frontier LLMs reach at most 30.4% on research math challenges and no model exceeds 50% on refusal for ill-posed problems.
-
Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs
Soohak is a new 439-problem mathematician-authored benchmark showing frontier LLMs reach only 30% on research math and fail to exceed 50% on refusing ill-posed questions.
-
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning task...
-
Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning
LLMs exhibit myopic planning in four-in-a-row: move choices are best explained by shallow nodes in reasoning traces, not the deep lookahead they generate, unlike humans where depth drives performance.
-
Crafting Reversible SFT Behaviors in Large Language Models
LCDD creates sparse carriers for SFT behaviors that SFT-Eraser can reverse, with ablations showing the sparse structure enables causal control.
-
LLM Translation of Compiler Intermediate Representation
IRIS-14B is the first LLM trained explicitly for GIMPLE-to-LLVM IR translation and outperforms much larger models by up to 44 percentage points on real-world C code.
-
Bringing Order to Asynchronous SGD: Towards Optimality under Data-Dependent Delays with Momentum
Momentum-based async SGD achieves optimal convergence rates for data-dependent delays without biasing updates toward simpler samples.
-
Characterizing the Expressivity of Local Attention in Transformers
Local attention strictly enlarges the class of regular languages recognizable by fixed-precision transformers by adding a second past operator in linear temporal logic, with global and local attention being expressive...
-
Fisher Markets with Approximately Optimal Bundles and the Need for a PCP Theorem for PPAD
PPAD-hardness for constant-δ approximate competitive equilibria in SPLC Fisher markets requires the PCP-for-PPAD conjecture, which the paper proves is necessary to establish such hardness.
-
Fisher Markets with Approximately Optimal Bundles and the Need for a PCP Theorem for PPAD
Establishes PPAD-hardness for (1-δ)-optimal bundle competitive equilibria in SPLC Fisher markets assuming PCP-for-PPAD conjecture, and proves the conjecture is required for hardness in this market class.
-
ArgBench: Benchmarking LLMs on Computational Argumentation Tasks
ArgBench unifies 33 existing datasets into a standardized benchmark for testing LLMs across 46 argumentation tasks and analyzes the impact of prompting techniques and model factors on performance.
-
MedPRMBench: A Fine-grained Benchmark for Process Reward Models in Medical Reasoning
MedPRMBench is the first fine-grained benchmark for process reward models in medical reasoning, featuring 6500 questions, 13000 chains, 113910 step labels, and a baseline that improves downstream QA accuracy by 3.2-6....
-
neuralCAD-Edit: An Expert Benchmark for Multimodal-Instructed 3D CAD Model Editing
neuralCAD-Edit benchmark shows even the best foundation model (GPT 5.2) scores 53% lower than human CAD experts in acceptance trials for multimodal-instructed 3D model edits.
-
Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation
Lightning OPD enforces teacher consistency by precomputing log-probabilities over SFT rollouts, matching standard OPD performance with bounded gradient discrepancy and achieving 4x speedup on math and code reasoning tasks.
-
Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models
Large language models display the identifiable victim effect at roughly twice the human baseline, strongly amplified by instruction tuning and chain-of-thought prompting but inverted by reasoning-specialized models.
-
NovBench: Evaluating Large Language Models on Academic Paper Novelty Assessment
NovBench is the first large-scale benchmark with 1,684 expert-annotated pairs to evaluate LLMs on assessing academic paper novelty via a four-dimensional framework of Relevance, Correctness, Coverage, and Clarity.
-
SARL: Label-Free Reinforcement Learning by Rewarding Reasoning Topology
SARL rewards reasoning topology to improve label-free RL, outperforming baselines with gains up to 44.7% on math and 34.6% on open-ended tasks while maintaining more stable training.
-
Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages
Derives an exact unbiased policy gradient for RL post-training of diffusion LLMs via entropy-guided step selection and one-step denoising rewards, achieving state-of-the-art results on coding and logical reasoning benchmarks.
-
MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs
MVI-Bench supplies the first taxonomy and dataset focused on misleading visual inputs to measure LVLM robustness, with tests on 18 models revealing clear weaknesses.
-
MediQAl: A French Medical Question Answering Dataset for Knowledge and Reasoning Evaluation
MediQAl is a new French medical QA benchmark with 32k exam-sourced questions in three formats and cognitive labels, evaluated on 14 LLMs to reveal gaps between factual recall and reasoning performance.
-
Flow-GRPO: Training Flow Matching Models via Online RL
Flow-GRPO is the first online RL method for flow matching models, raising GenEval accuracy from 63% to 95% and text-rendering accuracy from 59% to 92% with little reward hacking.
-
DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning
DeepMath-103K is a new 103K-problem mathematical dataset with high difficulty, rigorous decontamination, and verifiable answers to support RL training of language-model reasoning.
-
Large Language Diffusion Models
LLaDA is a scalable diffusion-based language model that matches autoregressive LLMs like LLaMA3 8B on tasks and surpasses GPT-4o on reversal poem completion.
-
DecompRL: Solving Harder Problems by Learning Modular Code Generation
DecompRL is an RL method that learns modular code decomposition for LLMs, enabling exponential candidate generation via recombination to solve harder coding problems with lower GPU cost.
-
DRIFTLENS: Measuring Memory-Induced Reasoning Drift in Personalized Language Models
DRIFTLENS quantifies memory-induced reasoning drift in personalized LLMs, finding medium-to-large effects across four models and ten user attributes that post-training only partly reduces.
-
Purified OPSD: On-Policy Self-Distillation Without Losing How to Think
Purified OPSD subtracts a reference-only teacher's signal from standard OPSD supervision and applies PMI to create a cleaner distillation target, yielding gains on long-CoT models while preserving epistemic behavior.
-
Evidence-State Rewards for Long-Context Reasoning
Maven is an RL method using answer-conditioned evidence-state values to assign rewards to add, link, and drop actions on evidence memory, outperforming outcome-only baselines on LongBench v2, LongReason, and RULER.
-
OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets
OpenSafeIntent benchmark shows models fail to calibrate safety across intent shifts in matched dual-use prompts, indicating current evaluations are insufficient.
-
Regression Accumulation in Multi-Turn LLM Programming Conversations
Regression accumulation affects 40-73% of 8-turn LLM coding tasks on extended HumanEval+/MBPP+ benchmarks, with verification gates improving final-turn pass rates on prior tests.
-
Measuring the Gap Between Human and LLM Research Ideas
LLM-generated research ideas cluster more around bridge-like opportunities and synthesis methods than the broader distribution seen in human papers.
-
Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning
P2R decouples perception from reasoning in VLMs via a two-stage process and PRA-GRPO alternating RL training, reporting gains such as 93.2% on V-Star for the 4B model over its Qwen3-VL backbone.
-
Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
Multi-agent LLMs generate and verify 14,073 deterministic reaction rules from 665,901 patents, enabling 97.7% classification of unseen reactions with finer resolution than fixed proprietary systems.
-
Know When to Stop: Segment-Level Credit Assignment for Reducing Overthinking
DASH assigns segment-level credit in reasoning traces using drift toward ground-truth answers, yielding 50.8% accuracy on AIME25 versus 45.4% for GRPO while reducing overthinking behaviors.
-
Verifiable Rewards for Calibrated Probabilistic Forecasting
A verifiable empirical win rate reward combined with gradient masking enables RL training of a 7B model to reach betting-market calibration on NFL win probabilities using only outcome data.
-
GRPO, Dr. GRPO, and DAPO Are Three Operations on One Number: The Group-Standard-Deviation Identity
GRPO, Dr. GRPO, and DAPO are three settings of one dial on the group standard deviation of binary rewards, unified by the group-standard-deviation identity where disagreement equals update magnitude.
-
Morphing into Hybrid Attention Models
FlashMorph formulates hybrid layer selection as budget-constrained optimization, trains per-layer gates on synthetic retrieval data with linearization regularization, then discretizes and distills to produce efficient...
-
When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon
Online IL overcomes an information-theoretic bottleneck that offline IL faces in non-realizable settings even at horizon 1, under a new structural characterization of reward-relative misspecification.
-
CORTEX: High-Quality Cross-Domain Organization of Web-Scale Corpora through Ontological Corpus Graph
Cortex uses an Ontological Corpus Graph to structure web-scale corpora, creating a refined 24.14B-token corpus and a new benchmark validated on eight LLMs.
-
On Test-Time Scaling for Vision-Language Models
Small well-performing LVLMs gain the most from test-time scaling with up to 30% improvements that can match or exceed larger models, while visual information is used mainly early in reasoning chains.
-
On Test-Time Scaling for Vision-Language Models
Small well-performing LVLMs gain the largest benefits from test-time scaling (up to ~30% improvement), often matching or exceeding larger models, while visual tokens contribute mainly early in the reasoning chain.
-
Personalizing MLLMs via Reinforced Multimodal Reference Game
RRG trains MLLMs via a reinforced multimodal reference game with contrastive rewards on hard positives and negatives to produce accurate, discriminative concept descriptions, achieving SOTA on personalization benchmarks.
-
Tandem Reinforcement Learning with Verifiable Rewards
TRL extends tandem training to RLVR pipelines, matching GRPO solo reasoning on Qwen3-4B math tasks while improving handoff robustness, reducing distributional drift, and increasing CoT legibility for the junior.
-
TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RL
TempAct introduces a planner-executor RL framework with hierarchical group exploration and rewards to improve temporal consistency in autoregressive video diffusion models.
-
LLM-Based Examination of Eligibility Criteria from Securities Prospectuses at the German Central Bank
LLMs are applied in a generative pipeline for extracting, normalizing, and interpreting eligibility criteria from securities prospectuses, achieving up to 91% precision in document-level decisions with a conservative bias.
-
FeVOS: Foresight Expression Video Object Segmentation
Introduces the FeVOS task, a 968-clip dataset with foresight expressions, and an MLLM model FeVOS-R1 trained via SFT then RL that reports SOTA on the new task plus generalization to prior RVOS benchmarks.
-
CrypFormBench: Benchmarking Formal Analysis Capability of Large Language Models for Cryptographic Schemes
CrypFormBench is a new benchmark jointly covering symbolic and computational security to evaluate LLMs on five formal analysis capabilities, with results showing top model Claude-3.5 scores 48.7/100 and most models st...
-
Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing
Local Branch Routing (LBR) is a token-level framework for test-time scaling in language models that uses local branch hidden states for routing and supports end-to-end RL, showing gains in Pass@1 and Pass@32 on math r...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.